Extend NanoBEIREvaluator to support custom NanoBEIR datasets

[milistu]

Description

This PR extends NanoBEIREvaluator to support custom HuggingFace NanoBEIR dataset paths, enabling evaluation on alternative NanoBEIR collections like:

• NanoBEIR with BM25 rankings
• Translated versions (e.g., Serbian)

Motivation

This feature was developed while creating an embedding model for the Serbian language. During this process, I noticed that under-resourced languages often lack evaluation benchmarks, making it difficult for practitioners to compare and select models for their applications.

By enabling NanoBEIREvaluator to accept custom dataset paths, researchers and developers can now:

• Evaluate models on translated NanoBEIR datasets
• Create and share benchmarks for their own languages
• Compare multilingual models on localised evaluation sets

I plan to publish my findings and a guide on translation and fine-tuning embedding models for low-resource languages in the near future.

Changes

NanoBEIREvaluator

• Accept custom HuggingFace dataset paths (e.g., "sentence-transformers/NanoClimateFEVER-bm25")
• Support "relevance" subset in addition to "qrels" for newer dataset formats
• Support positive-corpus-ids (list) in addition to corpus-id (single value)
• Validate custom paths follow NanoBEIR naming convention

Tests

• Updated existing validation test for the new error message
• Added tests for custom path validation

Usage Example

# Predefined datasets (existing behaviour)
evaluator = NanoBEIREvaluator(dataset_names=["climatefever", "msmarco"])

# Custom HuggingFace paths (new feature)
evaluator = NanoBEIREvaluator(
    dataset_names=["sentence-transformers/NanoClimateFEVER-bm25"]
)